Harnessing unsupervised learning for retrieving CAD assembly models from public datasets

Retrieving assembly models from public datasets can yield enriched outcomes and broaden the spectrum of insights. However, public datasets often present unique challenges, such as variance in quality and granularity of assembly models, lack of standardized methods for organizing and labeling, which hinder efficient and accurate retrieval. To address these issues, this paper presents a robust two-step retrieval method tailored for CAD assembly models from public datasets. The first phase utilizes hierarchical clustering in an unsupervised learning framework to systematically organize CAD assembly models. Each assembly model is represented by a feature vector that encapsulates geometrical and topological features derived from its Boundary Representation (B-rep), and reflects hierarchical relationships among parts and components. These feature vectors serve as the basis for systematic indexing via hierarchical clustering, grouping models based on similarity measurement. Each cluster’s centroid, representing the collective feature vector, facilitates efficient and targeted retrieval. In the second phase, the query model is directly compared to cluster centroids, enabling rapid identification of similar assembly collections. To enhance precision within identified clusters, we introduce a fine-grained retrieval technique that integrates Optimal Subsequence Bijection (OSB) with Maximum Mean Discrepancy (MMD). Evaluations on a heterogeneous dataset demonstrate that our method not only streamlines dataset organization but also effectively addresses quality variations, significantly improving retrieval efficiency across extensive collections.